[ecf6b46] | 1 | // |
---|
| 2 | // Cforall Version 1.0.0 Copyright (C) 2020 University of Waterloo |
---|
| 3 | // |
---|
| 4 | // The contents of this file are covered under the licence agreement in the |
---|
| 5 | // file "LICENCE" distributed with Cforall. |
---|
| 6 | // |
---|
| 7 | // io.cfa -- |
---|
| 8 | // |
---|
| 9 | // Author : Thierry Delisle |
---|
| 10 | // Created On : Thu Apr 23 17:31:00 2020 |
---|
| 11 | // Last Modified By : |
---|
| 12 | // Last Modified On : |
---|
| 13 | // Update Count : |
---|
| 14 | // |
---|
| 15 | |
---|
[3e2b9c9] | 16 | #define __cforall_thread__ |
---|
| 17 | |
---|
[20ab637] | 18 | #if defined(__CFA_DEBUG__) |
---|
| 19 | // #define __CFA_DEBUG_PRINT_IO__ |
---|
[e660761] | 20 | // #define __CFA_DEBUG_PRINT_IO_CORE__ |
---|
[20ab637] | 21 | #endif |
---|
[4069faad] | 22 | |
---|
[f6660520] | 23 | |
---|
[3e2b9c9] | 24 | #if defined(CFA_HAVE_LINUX_IO_URING_H) |
---|
[31bb2e1] | 25 | #define _GNU_SOURCE /* See feature_test_macros(7) */ |
---|
| 26 | #include <errno.h> |
---|
[3e2b9c9] | 27 | #include <signal.h> |
---|
[31bb2e1] | 28 | #include <stdint.h> |
---|
| 29 | #include <string.h> |
---|
| 30 | #include <unistd.h> |
---|
| 31 | |
---|
[92976d9] | 32 | extern "C" { |
---|
| 33 | #include <sys/syscall.h> |
---|
| 34 | |
---|
| 35 | #include <linux/io_uring.h> |
---|
| 36 | } |
---|
| 37 | |
---|
[3e2b9c9] | 38 | #include "stats.hfa" |
---|
| 39 | #include "kernel.hfa" |
---|
| 40 | #include "kernel/fwd.hfa" |
---|
| 41 | #include "io/types.hfa" |
---|
[185efe6] | 42 | |
---|
[426f60c] | 43 | static const char * opcodes[] = { |
---|
| 44 | "OP_NOP", |
---|
| 45 | "OP_READV", |
---|
| 46 | "OP_WRITEV", |
---|
| 47 | "OP_FSYNC", |
---|
| 48 | "OP_READ_FIXED", |
---|
| 49 | "OP_WRITE_FIXED", |
---|
| 50 | "OP_POLL_ADD", |
---|
| 51 | "OP_POLL_REMOVE", |
---|
| 52 | "OP_SYNC_FILE_RANGE", |
---|
| 53 | "OP_SENDMSG", |
---|
| 54 | "OP_RECVMSG", |
---|
| 55 | "OP_TIMEOUT", |
---|
| 56 | "OP_TIMEOUT_REMOVE", |
---|
| 57 | "OP_ACCEPT", |
---|
| 58 | "OP_ASYNC_CANCEL", |
---|
| 59 | "OP_LINK_TIMEOUT", |
---|
| 60 | "OP_CONNECT", |
---|
| 61 | "OP_FALLOCATE", |
---|
| 62 | "OP_OPENAT", |
---|
| 63 | "OP_CLOSE", |
---|
| 64 | "OP_FILES_UPDATE", |
---|
| 65 | "OP_STATX", |
---|
| 66 | "OP_READ", |
---|
| 67 | "OP_WRITE", |
---|
| 68 | "OP_FADVISE", |
---|
| 69 | "OP_MADVISE", |
---|
| 70 | "OP_SEND", |
---|
| 71 | "OP_RECV", |
---|
| 72 | "OP_OPENAT2", |
---|
| 73 | "OP_EPOLL_CTL", |
---|
| 74 | "OP_SPLICE", |
---|
| 75 | "OP_PROVIDE_BUFFERS", |
---|
| 76 | "OP_REMOVE_BUFFERS", |
---|
| 77 | "OP_TEE", |
---|
| 78 | "INVALID_OP" |
---|
| 79 | }; |
---|
| 80 | |
---|
[2fafe7e] | 81 | // returns true of acquired as leader or second leader |
---|
| 82 | static inline bool try_lock( __leaderlock_t & this ) { |
---|
| 83 | const uintptr_t thrd = 1z | (uintptr_t)active_thread(); |
---|
| 84 | bool block; |
---|
| 85 | disable_interrupts(); |
---|
| 86 | for() { |
---|
| 87 | struct $thread * expected = this.value; |
---|
| 88 | if( 1p != expected && 0p != expected ) { |
---|
| 89 | /* paranoid */ verify( thrd != (uintptr_t)expected ); // We better not already be the next leader |
---|
| 90 | enable_interrupts( __cfaabi_dbg_ctx ); |
---|
| 91 | return false; |
---|
| 92 | } |
---|
| 93 | struct $thread * desired; |
---|
| 94 | if( 0p == expected ) { |
---|
| 95 | // If the lock isn't locked acquire it, no need to block |
---|
| 96 | desired = 1p; |
---|
| 97 | block = false; |
---|
| 98 | } |
---|
| 99 | else { |
---|
| 100 | // If the lock is already locked try becomming the next leader |
---|
| 101 | desired = (struct $thread *)thrd; |
---|
| 102 | block = true; |
---|
| 103 | } |
---|
| 104 | if( __atomic_compare_exchange_n(&this.value, &expected, desired, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST) ) break; |
---|
| 105 | } |
---|
| 106 | if( block ) { |
---|
| 107 | enable_interrupts( __cfaabi_dbg_ctx ); |
---|
[e235429] | 108 | park(); |
---|
[2fafe7e] | 109 | disable_interrupts(); |
---|
| 110 | } |
---|
| 111 | return true; |
---|
| 112 | } |
---|
| 113 | |
---|
| 114 | static inline bool next( __leaderlock_t & this ) { |
---|
[8fc652e0] | 115 | /* paranoid */ verify( ! __preemption_enabled() ); |
---|
[2fafe7e] | 116 | struct $thread * nextt; |
---|
| 117 | for() { |
---|
| 118 | struct $thread * expected = this.value; |
---|
| 119 | /* paranoid */ verify( (1 & (uintptr_t)expected) == 1 ); // The lock better be locked |
---|
| 120 | |
---|
| 121 | struct $thread * desired; |
---|
| 122 | if( 1p == expected ) { |
---|
| 123 | // No next leader, just unlock |
---|
| 124 | desired = 0p; |
---|
| 125 | nextt = 0p; |
---|
| 126 | } |
---|
| 127 | else { |
---|
| 128 | // There is a next leader, remove but keep locked |
---|
| 129 | desired = 1p; |
---|
| 130 | nextt = (struct $thread *)(~1z & (uintptr_t)expected); |
---|
| 131 | } |
---|
| 132 | if( __atomic_compare_exchange_n(&this.value, &expected, desired, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST) ) break; |
---|
| 133 | } |
---|
| 134 | |
---|
| 135 | if(nextt) { |
---|
[b4b63e8] | 136 | unpark( nextt ); |
---|
[2fafe7e] | 137 | enable_interrupts( __cfaabi_dbg_ctx ); |
---|
| 138 | return true; |
---|
| 139 | } |
---|
| 140 | enable_interrupts( __cfaabi_dbg_ctx ); |
---|
| 141 | return false; |
---|
| 142 | } |
---|
| 143 | |
---|
[92976d9] | 144 | //============================================================================================= |
---|
[3e2b9c9] | 145 | // I/O Syscall |
---|
[92976d9] | 146 | //============================================================================================= |
---|
[3e2b9c9] | 147 | static int __io_uring_enter( struct __io_data & ring, unsigned to_submit, bool get ) { |
---|
[20ab637] | 148 | bool need_sys_to_submit = false; |
---|
| 149 | bool need_sys_to_complete = false; |
---|
| 150 | unsigned flags = 0; |
---|
| 151 | |
---|
| 152 | TO_SUBMIT: |
---|
| 153 | if( to_submit > 0 ) { |
---|
| 154 | if( !(ring.ring_flags & IORING_SETUP_SQPOLL) ) { |
---|
| 155 | need_sys_to_submit = true; |
---|
| 156 | break TO_SUBMIT; |
---|
| 157 | } |
---|
| 158 | if( (*ring.submit_q.flags) & IORING_SQ_NEED_WAKEUP ) { |
---|
| 159 | need_sys_to_submit = true; |
---|
| 160 | flags |= IORING_ENTER_SQ_WAKEUP; |
---|
| 161 | } |
---|
| 162 | } |
---|
| 163 | |
---|
| 164 | if( get && !(ring.ring_flags & IORING_SETUP_SQPOLL) ) { |
---|
| 165 | flags |= IORING_ENTER_GETEVENTS; |
---|
| 166 | if( (ring.ring_flags & IORING_SETUP_IOPOLL) ) { |
---|
| 167 | need_sys_to_complete = true; |
---|
| 168 | } |
---|
| 169 | } |
---|
| 170 | |
---|
| 171 | int ret = 0; |
---|
| 172 | if( need_sys_to_submit || need_sys_to_complete ) { |
---|
[ece0e80] | 173 | __cfadbg_print_safe(io_core, "Kernel I/O : IO_URING enter %d %u %u\n", ring.fd, to_submit, flags); |
---|
[b982fb2] | 174 | ret = syscall( __NR_io_uring_enter, ring.fd, to_submit, 0, flags, (sigset_t *)0p, _NSIG / 8); |
---|
[20ab637] | 175 | if( ret < 0 ) { |
---|
| 176 | switch((int)errno) { |
---|
| 177 | case EAGAIN: |
---|
| 178 | case EINTR: |
---|
| 179 | ret = -1; |
---|
| 180 | break; |
---|
| 181 | default: |
---|
| 182 | abort( "KERNEL ERROR: IO_URING SYSCALL - (%d) %s\n", (int)errno, strerror(errno) ); |
---|
| 183 | } |
---|
| 184 | } |
---|
| 185 | } |
---|
| 186 | |
---|
| 187 | // Memory barrier |
---|
| 188 | __atomic_thread_fence( __ATOMIC_SEQ_CST ); |
---|
| 189 | return ret; |
---|
| 190 | } |
---|
| 191 | |
---|
[92976d9] | 192 | //============================================================================================= |
---|
| 193 | // I/O Polling |
---|
| 194 | //============================================================================================= |
---|
[1d5e4711] | 195 | static unsigned __collect_submitions( struct __io_data & ring ); |
---|
[4998155] | 196 | static __u32 __release_consumed_submission( struct __io_data & ring ); |
---|
[1d5e4711] | 197 | |
---|
[426f60c] | 198 | // Process a single completion message from the io_uring |
---|
| 199 | // This is NOT thread-safe |
---|
| 200 | static inline void process( volatile struct io_uring_cqe & cqe ) { |
---|
[c402739f] | 201 | struct io_future_t * future = (struct io_future_t *)(uintptr_t)cqe.user_data; |
---|
[fe9468e2] | 202 | __cfadbg_print_safe( io, "Kernel I/O : Syscall completed : cqe %p, result %d for %p\n", &cqe, cqe.res, future ); |
---|
[5751a56] | 203 | |
---|
[c402739f] | 204 | fulfil( *future, cqe.res ); |
---|
[5751a56] | 205 | } |
---|
| 206 | |
---|
[f00b26d4] | 207 | static [int, bool] __drain_io( & struct __io_data ring ) { |
---|
[8fc652e0] | 208 | /* paranoid */ verify( ! __preemption_enabled() ); |
---|
[e46c753] | 209 | |
---|
[5dadc9b] | 210 | unsigned to_submit = 0; |
---|
[f00b26d4] | 211 | if( ring.poller_submits ) { |
---|
[5dadc9b] | 212 | // If the poller thread also submits, then we need to aggregate the submissions which are ready |
---|
[e46c753] | 213 | to_submit = __collect_submitions( ring ); |
---|
[5dadc9b] | 214 | } |
---|
| 215 | |
---|
[f00b26d4] | 216 | int ret = __io_uring_enter(ring, to_submit, true); |
---|
[20ab637] | 217 | if( ret < 0 ) { |
---|
| 218 | return [0, true]; |
---|
| 219 | } |
---|
[1d5e4711] | 220 | |
---|
[20ab637] | 221 | // update statistics |
---|
| 222 | if (to_submit > 0) { |
---|
[1d5e4711] | 223 | __STATS__( true, |
---|
| 224 | if( to_submit > 0 ) { |
---|
| 225 | io.submit_q.submit_avg.rdy += to_submit; |
---|
| 226 | io.submit_q.submit_avg.csm += ret; |
---|
| 227 | io.submit_q.submit_avg.cnt += 1; |
---|
| 228 | } |
---|
| 229 | ) |
---|
[6f121b8] | 230 | } |
---|
| 231 | |
---|
[426f60c] | 232 | __atomic_thread_fence( __ATOMIC_SEQ_CST ); |
---|
| 233 | |
---|
[20ab637] | 234 | // Release the consumed SQEs |
---|
| 235 | __release_consumed_submission( ring ); |
---|
[6f121b8] | 236 | |
---|
[d384787] | 237 | // Drain the queue |
---|
[92976d9] | 238 | unsigned head = *ring.completion_q.head; |
---|
[6f121b8] | 239 | unsigned tail = *ring.completion_q.tail; |
---|
[4998155] | 240 | const __u32 mask = *ring.completion_q.mask; |
---|
[6f121b8] | 241 | |
---|
[d384787] | 242 | // Nothing was new return 0 |
---|
| 243 | if (head == tail) { |
---|
[e46c753] | 244 | return [0, to_submit > 0]; |
---|
[d384787] | 245 | } |
---|
[92976d9] | 246 | |
---|
[4998155] | 247 | __u32 count = tail - head; |
---|
[1d5e4711] | 248 | /* paranoid */ verify( count != 0 ); |
---|
[d384787] | 249 | for(i; count) { |
---|
[6f121b8] | 250 | unsigned idx = (head + i) & mask; |
---|
[426f60c] | 251 | volatile struct io_uring_cqe & cqe = ring.completion_q.cqes[idx]; |
---|
[92976d9] | 252 | |
---|
[d384787] | 253 | /* paranoid */ verify(&cqe); |
---|
[92976d9] | 254 | |
---|
[f00b26d4] | 255 | process( cqe ); |
---|
[d384787] | 256 | } |
---|
[2d8f7b0] | 257 | |
---|
[92976d9] | 258 | // Mark to the kernel that the cqe has been seen |
---|
| 259 | // Ensure that the kernel only sees the new value of the head index after the CQEs have been read. |
---|
[426f60c] | 260 | __atomic_fetch_add( ring.completion_q.head, count, __ATOMIC_SEQ_CST ); |
---|
[92976d9] | 261 | |
---|
[5dadc9b] | 262 | return [count, count > 0 || to_submit > 0]; |
---|
[92976d9] | 263 | } |
---|
| 264 | |
---|
[f00b26d4] | 265 | void main( $io_ctx_thread & this ) { |
---|
[d48b174] | 266 | __ioctx_register( this ); |
---|
[1539bbd] | 267 | |
---|
[426f60c] | 268 | __cfadbg_print_safe(io_core, "Kernel I/O : IO poller %d (%p) ready\n", this.ring->fd, &this); |
---|
[1539bbd] | 269 | |
---|
[ece0e80] | 270 | const int reset_cnt = 5; |
---|
| 271 | int reset = reset_cnt; |
---|
[61dd73d] | 272 | // Then loop until we need to start |
---|
[ece0e80] | 273 | LOOP: |
---|
[f00b26d4] | 274 | while(!__atomic_load_n(&this.done, __ATOMIC_SEQ_CST)) { |
---|
[61dd73d] | 275 | // Drain the io |
---|
[5dadc9b] | 276 | int count; |
---|
| 277 | bool again; |
---|
[13c5e19] | 278 | disable_interrupts(); |
---|
[f00b26d4] | 279 | [count, again] = __drain_io( *this.ring ); |
---|
[5dadc9b] | 280 | |
---|
[ece0e80] | 281 | if(!again) reset--; |
---|
[3c039b0] | 282 | |
---|
[13c5e19] | 283 | // Update statistics |
---|
[47746a2] | 284 | __STATS__( true, |
---|
| 285 | io.complete_q.completed_avg.val += count; |
---|
[dcb5f8d] | 286 | io.complete_q.completed_avg.cnt += 1; |
---|
[47746a2] | 287 | ) |
---|
[13c5e19] | 288 | enable_interrupts( __cfaabi_dbg_ctx ); |
---|
[3c039b0] | 289 | |
---|
[5dadc9b] | 290 | // If we got something, just yield and check again |
---|
[ece0e80] | 291 | if(reset > 1) { |
---|
[61dd73d] | 292 | yield(); |
---|
[ece0e80] | 293 | continue LOOP; |
---|
[61dd73d] | 294 | } |
---|
[ece0e80] | 295 | |
---|
[426f60c] | 296 | // We alread failed to find completed entries a few time. |
---|
[ece0e80] | 297 | if(reset == 1) { |
---|
| 298 | // Rearm the context so it can block |
---|
| 299 | // but don't block right away |
---|
| 300 | // we need to retry one last time in case |
---|
| 301 | // something completed *just now* |
---|
[d48b174] | 302 | __ioctx_prepare_block( this ); |
---|
[ece0e80] | 303 | continue LOOP; |
---|
| 304 | } |
---|
| 305 | |
---|
[93526ef] | 306 | __STATS__( false, |
---|
[dcb5f8d] | 307 | io.complete_q.blocks += 1; |
---|
| 308 | ) |
---|
[426f60c] | 309 | __cfadbg_print_safe(io_core, "Kernel I/O : Parking io poller %d (%p)\n", this.ring->fd, &this); |
---|
[5dadc9b] | 310 | |
---|
[3e2b9c9] | 311 | // block this thread |
---|
[f00b26d4] | 312 | wait( this.sem ); |
---|
[ece0e80] | 313 | |
---|
| 314 | // restore counter |
---|
| 315 | reset = reset_cnt; |
---|
[f6660520] | 316 | } |
---|
[61dd73d] | 317 | |
---|
[426f60c] | 318 | __cfadbg_print_safe(io_core, "Kernel I/O : Fast poller %d (%p) stopping\n", this.ring->fd, &this); |
---|
[61dd73d] | 319 | } |
---|
[f6660520] | 320 | |
---|
[92976d9] | 321 | //============================================================================================= |
---|
| 322 | // I/O Submissions |
---|
| 323 | //============================================================================================= |
---|
| 324 | |
---|
[2d8f7b0] | 325 | // Submition steps : |
---|
[e46c753] | 326 | // 1 - Allocate a queue entry. The ring already has memory for all entries but only the ones |
---|
[2d8f7b0] | 327 | // listed in sq.array are visible by the kernel. For those not listed, the kernel does not |
---|
| 328 | // offer any assurance that an entry is not being filled by multiple flags. Therefore, we |
---|
| 329 | // need to write an allocator that allows allocating concurrently. |
---|
| 330 | // |
---|
[e46c753] | 331 | // 2 - Actually fill the submit entry, this is the only simple and straightforward step. |
---|
[2d8f7b0] | 332 | // |
---|
[e46c753] | 333 | // 3 - Append the entry index to the array and adjust the tail accordingly. This operation |
---|
[2d8f7b0] | 334 | // needs to arrive to two concensus at the same time: |
---|
| 335 | // A - The order in which entries are listed in the array: no two threads must pick the |
---|
| 336 | // same index for their entries |
---|
| 337 | // B - When can the tail be update for the kernel. EVERY entries in the array between |
---|
| 338 | // head and tail must be fully filled and shouldn't ever be touched again. |
---|
| 339 | // |
---|
| 340 | |
---|
[426f60c] | 341 | // Allocate an submit queue entry. |
---|
| 342 | // The kernel cannot see these entries until they are submitted, but other threads must be |
---|
| 343 | // able to see which entries can be used and which are already un used by an other thread |
---|
| 344 | // for convenience, return both the index and the pointer to the sqe |
---|
| 345 | // sqe == &sqes[idx] |
---|
| 346 | [* volatile struct io_uring_sqe, __u32] __submit_alloc( struct __io_data & ring, __u64 data ) { |
---|
[e46c753] | 347 | /* paranoid */ verify( data != 0 ); |
---|
[13c5e19] | 348 | |
---|
[6f121b8] | 349 | // Prepare the data we need |
---|
| 350 | __attribute((unused)) int len = 0; |
---|
| 351 | __attribute((unused)) int block = 0; |
---|
[4998155] | 352 | __u32 cnt = *ring.submit_q.num; |
---|
| 353 | __u32 mask = *ring.submit_q.mask; |
---|
[8ae4165] | 354 | |
---|
[fe9468e2] | 355 | __u32 off = thread_rand(); |
---|
[6f121b8] | 356 | |
---|
| 357 | // Loop around looking for an available spot |
---|
[13c5e19] | 358 | for() { |
---|
[6f121b8] | 359 | // Look through the list starting at some offset |
---|
| 360 | for(i; cnt) { |
---|
[426f60c] | 361 | __u64 expected = 3; |
---|
| 362 | __u32 idx = (i + off) & mask; // Get an index from a random |
---|
| 363 | volatile struct io_uring_sqe * sqe = &ring.submit_q.sqes[idx]; |
---|
[4998155] | 364 | volatile __u64 * udata = &sqe->user_data; |
---|
[6f121b8] | 365 | |
---|
[426f60c] | 366 | // Allocate the entry by CASing the user_data field from 0 to the future address |
---|
[6f121b8] | 367 | if( *udata == expected && |
---|
| 368 | __atomic_compare_exchange_n( udata, &expected, data, true, __ATOMIC_SEQ_CST, __ATOMIC_RELAXED ) ) |
---|
| 369 | { |
---|
| 370 | // update statistics |
---|
[47746a2] | 371 | __STATS__( false, |
---|
| 372 | io.submit_q.alloc_avg.val += len; |
---|
| 373 | io.submit_q.alloc_avg.block += block; |
---|
| 374 | io.submit_q.alloc_avg.cnt += 1; |
---|
| 375 | ) |
---|
[6f121b8] | 376 | |
---|
[426f60c] | 377 | // debug log |
---|
[ece0e80] | 378 | __cfadbg_print_safe( io, "Kernel I/O : allocated [%p, %u] for %p (%p)\n", sqe, idx, active_thread(), (void*)data ); |
---|
[13c5e19] | 379 | |
---|
[6f121b8] | 380 | // Success return the data |
---|
[426f60c] | 381 | sqe->opcode = 0; |
---|
| 382 | sqe->flags = 0; |
---|
| 383 | sqe->ioprio = 0; |
---|
| 384 | sqe->fd = 0; |
---|
| 385 | sqe->off = 0; |
---|
| 386 | sqe->addr = 0; |
---|
| 387 | sqe->len = 0; |
---|
| 388 | sqe->accept_flags = 0; |
---|
| 389 | sqe->__pad2[0] = 0; |
---|
| 390 | sqe->__pad2[1] = 0; |
---|
| 391 | sqe->__pad2[2] = 0; |
---|
[6f121b8] | 392 | return [sqe, idx]; |
---|
| 393 | } |
---|
| 394 | verify(expected != data); |
---|
[2489d31] | 395 | |
---|
[426f60c] | 396 | // This one was used |
---|
[6f121b8] | 397 | len ++; |
---|
| 398 | } |
---|
[2489d31] | 399 | |
---|
[6f121b8] | 400 | block++; |
---|
[426f60c] | 401 | |
---|
| 402 | abort( "Kernel I/O : all submit queue entries used, yielding\n" ); |
---|
| 403 | |
---|
[6f121b8] | 404 | yield(); |
---|
| 405 | } |
---|
[2489d31] | 406 | } |
---|
| 407 | |
---|
[4998155] | 408 | static inline __u32 __submit_to_ready_array( struct __io_data & ring, __u32 idx, const __u32 mask ) { |
---|
[df40a56] | 409 | /* paranoid */ verify( idx <= mask ); |
---|
| 410 | /* paranoid */ verify( idx != -1ul32 ); |
---|
| 411 | |
---|
| 412 | // We need to find a spot in the ready array |
---|
| 413 | __attribute((unused)) int len = 0; |
---|
| 414 | __attribute((unused)) int block = 0; |
---|
[4998155] | 415 | __u32 ready_mask = ring.submit_q.ready_cnt - 1; |
---|
[df40a56] | 416 | |
---|
[fe9468e2] | 417 | __u32 off = thread_rand(); |
---|
[df40a56] | 418 | |
---|
[4998155] | 419 | __u32 picked; |
---|
[df40a56] | 420 | LOOKING: for() { |
---|
| 421 | for(i; ring.submit_q.ready_cnt) { |
---|
| 422 | picked = (i + off) & ready_mask; |
---|
[4998155] | 423 | __u32 expected = -1ul32; |
---|
[df40a56] | 424 | if( __atomic_compare_exchange_n( &ring.submit_q.ready[picked], &expected, idx, true, __ATOMIC_SEQ_CST, __ATOMIC_RELAXED ) ) { |
---|
| 425 | break LOOKING; |
---|
| 426 | } |
---|
| 427 | verify(expected != idx); |
---|
| 428 | |
---|
| 429 | len ++; |
---|
| 430 | } |
---|
| 431 | |
---|
| 432 | block++; |
---|
[2fafe7e] | 433 | |
---|
| 434 | __u32 released = __release_consumed_submission( ring ); |
---|
| 435 | if( released == 0 ) { |
---|
[34b61882] | 436 | yield(); |
---|
| 437 | } |
---|
[df40a56] | 438 | } |
---|
| 439 | |
---|
| 440 | // update statistics |
---|
[47746a2] | 441 | __STATS__( false, |
---|
| 442 | io.submit_q.look_avg.val += len; |
---|
| 443 | io.submit_q.look_avg.block += block; |
---|
| 444 | io.submit_q.look_avg.cnt += 1; |
---|
| 445 | ) |
---|
[df40a56] | 446 | |
---|
| 447 | return picked; |
---|
| 448 | } |
---|
| 449 | |
---|
[4998155] | 450 | void __submit( struct io_context * ctx, __u32 idx ) __attribute__((nonnull (1))) { |
---|
[f00b26d4] | 451 | __io_data & ring = *ctx->thrd.ring; |
---|
[426f60c] | 452 | |
---|
| 453 | { |
---|
| 454 | __attribute__((unused)) volatile struct io_uring_sqe * sqe = &ring.submit_q.sqes[idx]; |
---|
| 455 | __cfadbg_print_safe( io, |
---|
| 456 | "Kernel I/O : submitting %u (%p) for %p\n" |
---|
| 457 | " data: %p\n" |
---|
| 458 | " opcode: %s\n" |
---|
| 459 | " fd: %d\n" |
---|
| 460 | " flags: %d\n" |
---|
| 461 | " prio: %d\n" |
---|
| 462 | " off: %p\n" |
---|
| 463 | " addr: %p\n" |
---|
| 464 | " len: %d\n" |
---|
| 465 | " other flags: %d\n" |
---|
| 466 | " splice fd: %d\n" |
---|
| 467 | " pad[0]: %llu\n" |
---|
| 468 | " pad[1]: %llu\n" |
---|
| 469 | " pad[2]: %llu\n", |
---|
| 470 | idx, sqe, |
---|
| 471 | active_thread(), |
---|
| 472 | (void*)sqe->user_data, |
---|
| 473 | opcodes[sqe->opcode], |
---|
| 474 | sqe->fd, |
---|
| 475 | sqe->flags, |
---|
| 476 | sqe->ioprio, |
---|
| 477 | sqe->off, |
---|
| 478 | sqe->addr, |
---|
| 479 | sqe->len, |
---|
| 480 | sqe->accept_flags, |
---|
| 481 | sqe->splice_fd_in, |
---|
| 482 | sqe->__pad2[0], |
---|
| 483 | sqe->__pad2[1], |
---|
| 484 | sqe->__pad2[2] |
---|
| 485 | ); |
---|
| 486 | } |
---|
| 487 | |
---|
| 488 | |
---|
[5dadc9b] | 489 | // Get now the data we definetely need |
---|
[4998155] | 490 | volatile __u32 * const tail = ring.submit_q.tail; |
---|
| 491 | const __u32 mask = *ring.submit_q.mask; |
---|
[2489d31] | 492 | |
---|
[5dadc9b] | 493 | // There are 2 submission schemes, check which one we are using |
---|
[f00b26d4] | 494 | if( ring.poller_submits ) { |
---|
[5dadc9b] | 495 | // If the poller thread submits, then we just need to add this to the ready array |
---|
[df40a56] | 496 | __submit_to_ready_array( ring, idx, mask ); |
---|
[5dadc9b] | 497 | |
---|
[f00b26d4] | 498 | post( ctx->thrd.sem ); |
---|
[5dadc9b] | 499 | |
---|
[dd4e2d7] | 500 | __cfadbg_print_safe( io, "Kernel I/O : Added %u to ready for %p\n", idx, active_thread() ); |
---|
[2d8f7b0] | 501 | } |
---|
[f00b26d4] | 502 | else if( ring.eager_submits ) { |
---|
[4998155] | 503 | __u32 picked = __submit_to_ready_array( ring, idx, mask ); |
---|
[e46c753] | 504 | |
---|
[2fafe7e] | 505 | #if defined(LEADER_LOCK) |
---|
| 506 | if( !try_lock(ring.submit_q.submit_lock) ) { |
---|
[47746a2] | 507 | __STATS__( false, |
---|
| 508 | io.submit_q.helped += 1; |
---|
| 509 | ) |
---|
[e46c753] | 510 | return; |
---|
| 511 | } |
---|
[8fc652e0] | 512 | /* paranoid */ verify( ! __preemption_enabled() ); |
---|
[2fafe7e] | 513 | __STATS__( true, |
---|
| 514 | io.submit_q.leader += 1; |
---|
| 515 | ) |
---|
| 516 | #else |
---|
| 517 | for() { |
---|
| 518 | yield(); |
---|
| 519 | |
---|
| 520 | if( try_lock(ring.submit_q.submit_lock __cfaabi_dbg_ctx2) ) { |
---|
| 521 | __STATS__( false, |
---|
| 522 | io.submit_q.leader += 1; |
---|
| 523 | ) |
---|
| 524 | break; |
---|
| 525 | } |
---|
| 526 | |
---|
| 527 | // If some one else collected our index, we are done |
---|
| 528 | #warning ABA problem |
---|
| 529 | if( ring.submit_q.ready[picked] != idx ) { |
---|
| 530 | __STATS__( false, |
---|
| 531 | io.submit_q.helped += 1; |
---|
| 532 | ) |
---|
| 533 | return; |
---|
| 534 | } |
---|
[e46c753] | 535 | |
---|
[47746a2] | 536 | __STATS__( false, |
---|
[2fafe7e] | 537 | io.submit_q.busy += 1; |
---|
[47746a2] | 538 | ) |
---|
[e46c753] | 539 | } |
---|
[2fafe7e] | 540 | #endif |
---|
[e46c753] | 541 | |
---|
| 542 | // We got the lock |
---|
[be36ec3] | 543 | // Collect the submissions |
---|
[e46c753] | 544 | unsigned to_submit = __collect_submitions( ring ); |
---|
| 545 | |
---|
[be36ec3] | 546 | // Actually submit |
---|
| 547 | int ret = __io_uring_enter( ring, to_submit, false ); |
---|
[309d814] | 548 | |
---|
[2fafe7e] | 549 | #if defined(LEADER_LOCK) |
---|
[8fc652e0] | 550 | /* paranoid */ verify( ! __preemption_enabled() ); |
---|
[2fafe7e] | 551 | next(ring.submit_q.submit_lock); |
---|
| 552 | #else |
---|
| 553 | unlock(ring.submit_q.submit_lock); |
---|
| 554 | #endif |
---|
[ece0e80] | 555 | if( ret < 0 ) { |
---|
| 556 | return; |
---|
| 557 | } |
---|
[e46c753] | 558 | |
---|
| 559 | // Release the consumed SQEs |
---|
[34b61882] | 560 | __release_consumed_submission( ring ); |
---|
[e46c753] | 561 | |
---|
| 562 | // update statistics |
---|
[be36ec3] | 563 | __STATS__( false, |
---|
[47746a2] | 564 | io.submit_q.submit_avg.rdy += to_submit; |
---|
| 565 | io.submit_q.submit_avg.csm += ret; |
---|
| 566 | io.submit_q.submit_avg.cnt += 1; |
---|
| 567 | ) |
---|
[ece0e80] | 568 | |
---|
| 569 | __cfadbg_print_safe( io, "Kernel I/O : submitted %u (among %u) for %p\n", idx, ret, active_thread() ); |
---|
[e46c753] | 570 | } |
---|
[426f60c] | 571 | else |
---|
| 572 | { |
---|
[5dadc9b] | 573 | // get mutual exclusion |
---|
[2fafe7e] | 574 | #if defined(LEADER_LOCK) |
---|
| 575 | while(!try_lock(ring.submit_q.submit_lock)); |
---|
| 576 | #else |
---|
| 577 | lock(ring.submit_q.submit_lock __cfaabi_dbg_ctx2); |
---|
| 578 | #endif |
---|
[2489d31] | 579 | |
---|
[426f60c] | 580 | /* paranoid */ verifyf( ring.submit_q.sqes[ idx ].user_data != 3ul64, |
---|
[20ab637] | 581 | /* paranoid */ "index %u already reclaimed\n" |
---|
| 582 | /* paranoid */ "head %u, prev %u, tail %u\n" |
---|
| 583 | /* paranoid */ "[-0: %u,-1: %u,-2: %u,-3: %u]\n", |
---|
| 584 | /* paranoid */ idx, |
---|
| 585 | /* paranoid */ *ring.submit_q.head, ring.submit_q.prev_head, *tail |
---|
| 586 | /* paranoid */ ,ring.submit_q.array[ ((*ring.submit_q.head) - 0) & (*ring.submit_q.mask) ] |
---|
| 587 | /* paranoid */ ,ring.submit_q.array[ ((*ring.submit_q.head) - 1) & (*ring.submit_q.mask) ] |
---|
| 588 | /* paranoid */ ,ring.submit_q.array[ ((*ring.submit_q.head) - 2) & (*ring.submit_q.mask) ] |
---|
| 589 | /* paranoid */ ,ring.submit_q.array[ ((*ring.submit_q.head) - 3) & (*ring.submit_q.mask) ] |
---|
| 590 | /* paranoid */ ); |
---|
| 591 | |
---|
[5dadc9b] | 592 | // Append to the list of ready entries |
---|
| 593 | |
---|
| 594 | /* paranoid */ verify( idx <= mask ); |
---|
[20ab637] | 595 | ring.submit_q.array[ (*tail) & mask ] = idx; |
---|
[5dadc9b] | 596 | __atomic_fetch_add(tail, 1ul32, __ATOMIC_SEQ_CST); |
---|
[d384787] | 597 | |
---|
[5dadc9b] | 598 | // Submit however, many entries need to be submitted |
---|
[f00b26d4] | 599 | int ret = __io_uring_enter( ring, 1, false ); |
---|
[5dadc9b] | 600 | if( ret < 0 ) { |
---|
| 601 | switch((int)errno) { |
---|
| 602 | default: |
---|
| 603 | abort( "KERNEL ERROR: IO_URING SUBMIT - %s\n", strerror(errno) ); |
---|
| 604 | } |
---|
| 605 | } |
---|
[d384787] | 606 | |
---|
[426f60c] | 607 | /* paranoid */ verify(ret == 1); |
---|
| 608 | |
---|
[5dadc9b] | 609 | // update statistics |
---|
[47746a2] | 610 | __STATS__( false, |
---|
| 611 | io.submit_q.submit_avg.csm += 1; |
---|
| 612 | io.submit_q.submit_avg.cnt += 1; |
---|
| 613 | ) |
---|
[5dadc9b] | 614 | |
---|
[426f60c] | 615 | { |
---|
| 616 | __attribute__((unused)) volatile __u32 * const head = ring.submit_q.head; |
---|
| 617 | __attribute__((unused)) __u32 last_idx = ring.submit_q.array[ ((*head) - 1) & mask ]; |
---|
| 618 | __attribute__((unused)) volatile struct io_uring_sqe * sqe = &ring.submit_q.sqes[last_idx]; |
---|
| 619 | |
---|
| 620 | __cfadbg_print_safe( io, |
---|
| 621 | "Kernel I/O : last submitted is %u (%p)\n" |
---|
| 622 | " data: %p\n" |
---|
| 623 | " opcode: %s\n" |
---|
| 624 | " fd: %d\n" |
---|
| 625 | " flags: %d\n" |
---|
| 626 | " prio: %d\n" |
---|
| 627 | " off: %p\n" |
---|
| 628 | " addr: %p\n" |
---|
| 629 | " len: %d\n" |
---|
| 630 | " other flags: %d\n" |
---|
| 631 | " splice fd: %d\n" |
---|
| 632 | " pad[0]: %llu\n" |
---|
| 633 | " pad[1]: %llu\n" |
---|
| 634 | " pad[2]: %llu\n", |
---|
| 635 | last_idx, sqe, |
---|
| 636 | (void*)sqe->user_data, |
---|
| 637 | opcodes[sqe->opcode], |
---|
| 638 | sqe->fd, |
---|
| 639 | sqe->flags, |
---|
| 640 | sqe->ioprio, |
---|
| 641 | sqe->off, |
---|
| 642 | sqe->addr, |
---|
| 643 | sqe->len, |
---|
| 644 | sqe->accept_flags, |
---|
| 645 | sqe->splice_fd_in, |
---|
| 646 | sqe->__pad2[0], |
---|
| 647 | sqe->__pad2[1], |
---|
| 648 | sqe->__pad2[2] |
---|
| 649 | ); |
---|
| 650 | } |
---|
| 651 | |
---|
| 652 | __atomic_thread_fence( __ATOMIC_SEQ_CST ); |
---|
[34b61882] | 653 | // Release the consumed SQEs |
---|
| 654 | __release_consumed_submission( ring ); |
---|
[426f60c] | 655 | // ring.submit_q.sqes[idx].user_data = 3ul64; |
---|
[7bfc849] | 656 | |
---|
[2fafe7e] | 657 | #if defined(LEADER_LOCK) |
---|
| 658 | next(ring.submit_q.submit_lock); |
---|
| 659 | #else |
---|
| 660 | unlock(ring.submit_q.submit_lock); |
---|
| 661 | #endif |
---|
[dd4e2d7] | 662 | |
---|
[426f60c] | 663 | __cfadbg_print_safe( io, "Kernel I/O : submitted %u for %p\n", idx, active_thread() ); |
---|
[5dadc9b] | 664 | } |
---|
[2489d31] | 665 | } |
---|
[e46c753] | 666 | |
---|
[d2b5d2d] | 667 | // #define PARTIAL_SUBMIT 32 |
---|
[426f60c] | 668 | |
---|
| 669 | // go through the list of submissions in the ready array and moved them into |
---|
| 670 | // the ring's submit queue |
---|
[e46c753] | 671 | static unsigned __collect_submitions( struct __io_data & ring ) { |
---|
| 672 | /* paranoid */ verify( ring.submit_q.ready != 0p ); |
---|
| 673 | /* paranoid */ verify( ring.submit_q.ready_cnt > 0 ); |
---|
| 674 | |
---|
| 675 | unsigned to_submit = 0; |
---|
[4998155] | 676 | __u32 tail = *ring.submit_q.tail; |
---|
| 677 | const __u32 mask = *ring.submit_q.mask; |
---|
[1095ccd] | 678 | #if defined(PARTIAL_SUBMIT) |
---|
| 679 | #if defined(LEADER_LOCK) |
---|
| 680 | #error PARTIAL_SUBMIT and LEADER_LOCK cannot co-exist |
---|
| 681 | #endif |
---|
| 682 | const __u32 cnt = ring.submit_q.ready_cnt > PARTIAL_SUBMIT ? PARTIAL_SUBMIT : ring.submit_q.ready_cnt; |
---|
| 683 | const __u32 offset = ring.submit_q.prev_ready; |
---|
| 684 | ring.submit_q.prev_ready += cnt; |
---|
| 685 | #else |
---|
| 686 | const __u32 cnt = ring.submit_q.ready_cnt; |
---|
| 687 | const __u32 offset = 0; |
---|
| 688 | #endif |
---|
[e46c753] | 689 | |
---|
| 690 | // Go through the list of ready submissions |
---|
[1095ccd] | 691 | for( c; cnt ) { |
---|
| 692 | __u32 i = (offset + c) % ring.submit_q.ready_cnt; |
---|
| 693 | |
---|
[e46c753] | 694 | // replace any submission with the sentinel, to consume it. |
---|
[4998155] | 695 | __u32 idx = __atomic_exchange_n( &ring.submit_q.ready[i], -1ul32, __ATOMIC_RELAXED); |
---|
[e46c753] | 696 | |
---|
| 697 | // If it was already the sentinel, then we are done |
---|
| 698 | if( idx == -1ul32 ) continue; |
---|
| 699 | |
---|
| 700 | // If we got a real submission, append it to the list |
---|
| 701 | ring.submit_q.array[ (tail + to_submit) & mask ] = idx & mask; |
---|
| 702 | to_submit++; |
---|
| 703 | } |
---|
| 704 | |
---|
| 705 | // Increment the tail based on how many we are ready to submit |
---|
| 706 | __atomic_fetch_add(ring.submit_q.tail, to_submit, __ATOMIC_SEQ_CST); |
---|
| 707 | |
---|
| 708 | return to_submit; |
---|
| 709 | } |
---|
[34b61882] | 710 | |
---|
[426f60c] | 711 | // Go through the ring's submit queue and release everything that has already been consumed |
---|
| 712 | // by io_uring |
---|
[4998155] | 713 | static __u32 __release_consumed_submission( struct __io_data & ring ) { |
---|
| 714 | const __u32 smask = *ring.submit_q.mask; |
---|
[732b406] | 715 | |
---|
[426f60c] | 716 | // We need to get the lock to copy the old head and new head |
---|
[732b406] | 717 | if( !try_lock(ring.submit_q.release_lock __cfaabi_dbg_ctx2) ) return 0; |
---|
[426f60c] | 718 | __attribute__((unused)) |
---|
| 719 | __u32 ctail = *ring.submit_q.tail; // get the current tail of the queue |
---|
| 720 | __u32 chead = *ring.submit_q.head; // get the current head of the queue |
---|
| 721 | __u32 phead = ring.submit_q.prev_head; // get the head the last time we were here |
---|
| 722 | ring.submit_q.prev_head = chead; // note up to were we processed |
---|
[732b406] | 723 | unlock(ring.submit_q.release_lock); |
---|
| 724 | |
---|
[426f60c] | 725 | // the 3 fields are organized like this diagram |
---|
| 726 | // except it's are ring |
---|
| 727 | // ---+--------+--------+---- |
---|
| 728 | // ---+--------+--------+---- |
---|
| 729 | // ^ ^ ^ |
---|
| 730 | // phead chead ctail |
---|
| 731 | |
---|
| 732 | // make sure ctail doesn't wrap around and reach phead |
---|
| 733 | /* paranoid */ verify( |
---|
| 734 | (ctail >= chead && chead >= phead) |
---|
| 735 | || (chead >= phead && phead >= ctail) |
---|
| 736 | || (phead >= ctail && ctail >= chead) |
---|
| 737 | ); |
---|
| 738 | |
---|
| 739 | // find the range we need to clear |
---|
[4998155] | 740 | __u32 count = chead - phead; |
---|
[426f60c] | 741 | |
---|
| 742 | // We acquired an previous-head/current-head range |
---|
| 743 | // go through the range and release the sqes |
---|
[34b61882] | 744 | for( i; count ) { |
---|
[4998155] | 745 | __u32 idx = ring.submit_q.array[ (phead + i) & smask ]; |
---|
[426f60c] | 746 | |
---|
| 747 | /* paranoid */ verify( 0 != ring.submit_q.sqes[ idx ].user_data ); |
---|
| 748 | ring.submit_q.sqes[ idx ].user_data = 3ul64; |
---|
[34b61882] | 749 | } |
---|
| 750 | return count; |
---|
| 751 | } |
---|
[47746a2] | 752 | #endif |
---|